
We gave 10 frontier AI models wallets with money. 7 got scammed at least once. The best wasted $25 of every $100. The worst wasted $71. Agent Money Benchmark tests whether AI agents can be trusted with real money. We evaluated 10 frontier models across two settings, using real USDC payments on @solana : → Financial Research: Agents buy live data from real providers to answer open-ended financial questions (prediction markets, wallet analysis, DeFi yields, etc.). → Adversarial Marketplace: Agents buy data from simulated sellers, some of whom cheat. The tasks test whether agents can avoid payment traps, stay within budget, and complete transactions without being exploited. Key results: → Give an agent $100 and the best tier wastes $25–38 of it. The weakest tier wastes $56–71. → 7 of 10 models paid a cheating seller at least once, and 6 paid twice for the same data. → Only 4 of 10 models made no money errors. → No model gets more than half of the research questions right. GPT-6.1 Sol led at 50%, at $0.468 per question. Explore the full results → agentmoneybench.com/



















